Hybrid Feature Learning Using Autoencoders for Early Prediction of Sepsis

The early prediction of sepsis is important for ICU patients, as the risk of mortality increases as the disease is left untreated. We hypothesize that there is a need to learn important feature representations, such as to extract salient information from sepsis data. In this paper, we propose an unsupervised method to learn spatial-temporal information from the data, through the use of two autoencoders. For the official 2019 PhysioNet Challenge, our team, Kent Ridge AI (ranked 77th), obtained a utility score of-0.164 on the full test set. Additionally, we report crossvalidation results and identify several issues which can potentially help to improve performance.